MLS-C01 Hard Questions — 25 Difficult Questions

Question 1

A data engineering team is designing a streaming data pipeline that ingests 10,000 events per second. Each event is 2 KB. The pipeline must process events with a latency of less than 1 second. The team is considering using Amazon Kinesis Data Streams with 10 shards. Which TWO additional configurations should the team implement to meet the latency requirement? (Choose two.)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Enable auto-scaling of shards based on throughput.. Correct options: D and E. Enhanced fan-out (E) provides dedicated read throughput of 2 MB/second per shard per consumer, eliminating contention and ensuring low latency (<1 second) even with multiple consumers. Auto-scaling shards (D) dynamically adjusts shard count based on incoming throughput, which is critical because the current 10 shards can only ingest 10 MB/sec (1 MB/sec write per shard) but the input is 20 MB/sec (10,000 events × 2 KB each). Option A (record aggregation) reduces PUT request count but does not affect read latency. Option B (Kinesis Firehose to S3) is for data delivery, not latency. Option C (increasing shards to 100) provides excess capacity but is costly and less efficient than auto-scaling.

Answer

Use record aggregation to reduce the number of records.

Answer

Configure Kinesis Data Firehose to deliver data to Amazon S3.

Answer

Increase the number of shards to 100.

Question 2

An ML team is deploying a model to a SageMaker endpoint for real-time inference. The model is large (2 GB) and requires GPU for low-latency inference. The team wants to minimize cost while maintaining a response time of under 200 ms. Which instance configuration and SageMaker feature would be best?

Accepted Answer

Use a GPU instance (ml.p3.2xlarge) with SageMaker Elastic Inference.. The best choice. By using a GPU instance (ml.p3.2xlarge) combined with SageMaker Elastic Inference, the team can attach a fractional GPU accelerator to the instance, providing the necessary GPU compute for low-latency inference while minimizing cost compared to using a full GPU instance alone. Option B (batch transform) is designed for offline inference on large datasets, not real-time. Option C (serverless inference) typically uses CPU instances and may incur cold starts, which can exceed the 200 ms response time requirement. Option D (multi-model endpoint) is optimized for hosting multiple models on a single instance, but does not directly address GPU acceleration or cost minimization for a single large model.

Answer

Use a batch transform job on GPU instances.

Answer

Use a serverless inference endpoint with a CPU instance.

Answer

Use a multi-model endpoint on a GPU instance.

Question 3

A company uses Amazon SageMaker to deploy a model for real-time predictions. The model is updated weekly. The company wants to ensure that the new model version is gradually rolled out to a small percentage of traffic before full deployment, and that it can be rolled back quickly if issues are detected. Which deployment strategy should be used?

Accepted Answer

Canary deployment using SageMaker endpoint variants. Amazon SageMaker endpoint variants support canary deployments, where you can shift a small percentage of traffic to a new model version (e.g., 5%) while the majority remains on the old version. This allows gradual rollout and immediate rollback by simply adjusting the traffic distribution weights or deleting the new variant, meeting the requirement for quick rollback without redeploying.

Answer

Blue/green deployment

Answer

A/B testing with a holdout group

Answer

Rolling deployment across multiple endpoints

Question 4

A team wants to automate the retraining of a model weekly using new data that arrives in S3. Which combination of services should they use?

Accepted Answer

Amazon SageMaker Pipelines and S3 events. Amazon SageMaker Pipelines natively supports automated retraining workflows triggered by S3 events. When new data arrives in S3, an event notification can invoke a Lambda function that starts a pipeline execution, which includes steps for data processing, training, evaluation, and model registration. This provides a fully managed, repeatable, and auditable MLOps pipeline without custom orchestration code.

Answer

AWS Lambda and S3 events

Answer

Amazon SageMaker Processing jobs

Answer

AWS Step Functions and AWS Glue

Question 5

A data engineer needs to set up a data lake on S3 that supports both batch and streaming ingestion. The data must be queryable by Athena, Redshift Spectrum, and EMR. Which TWO configurations are essential? (Choose two.)

Accepted Answer

Store data in columnar formats like Parquet or ORC.. Columnar formats like Parquet and ORC are optimized for analytical queries, reducing I/O by reading only the necessary columns. This is essential for Athena, Redshift Spectrum, and EMR, which all benefit from the efficient compression and predicate pushdown capabilities of these formats, enabling faster query performance and lower costs.

Answer

Enable S3 Select on the target buckets.

Answer

Enable S3 versioning on all buckets.

Answer

Set up Kinesis Data Firehose for streaming ingestion.

Question 6

A company is using Amazon SageMaker to build a custom model. The training job is failing with a 'ResourceLimitExceeded' error. Which TWO actions should the company take to resolve this issue?

Accepted Answer

Request a service quota increase for SageMaker training instances.. The 'ResourceLimitExceeded' error indicates that the AWS account has reached its service quota for SageMaker training instances. Requesting a service quota increase through the AWS Service Quotas console or API allows the account to launch additional instances of the required type, directly resolving the limit issue.

Answer

Use Amazon EFS for training data.

Answer

Reduce the size of the training dataset.

Answer

Use a smaller instance type.

Question 7

A data scientist is performing EDA on a dataset with 1,000 features and 10,000 rows. The target variable is binary. After checking for multicollinearity, the scientist finds many pairs of features with correlation > 0.95. Which action should be taken to prepare the data for modeling?

Accepted Answer

For each highly correlated pair, remove one feature based on domain knowledge or higher correlation with target.. When features are highly correlated (e.g., > 0.95), they introduce multicollinearity, which can destabilize coefficient estimates in linear models and reduce interpretability. Removing one feature from each correlated pair based on domain knowledge or its correlation with the target variable preserves predictive power while reducing redundancy. This approach is more targeted than PCA, which transforms features into uncorrelated components but sacrifices interpretability and may not align with the binary target.

Answer

Apply PCA to all features to decorrelate them.

Answer

Standardize all features using StandardScaler.

Answer

Randomly drop half of the correlated features.

Question 8

A data scientist is training a model using Amazon SageMaker. The training dataset is 500 GB and is stored in S3. The data scientist wants to use Pipe input mode to stream data directly from S3 to the training container. However, the training job fails with an error indicating that the container cannot read the data. What is the most likely cause?

Accepted Answer

The training algorithm does not support Pipe mode. Pipe mode in SageMaker streams data directly from S3 to the training container via a FIFO pipe, but the training algorithm must be specifically designed to read from a pipe (e.g., via stdin or a named pipe). If the algorithm expects random access to files or uses libraries that require seekable file handles (like many image-processing or custom Python scripts using `open()`), it will fail because a pipe is a sequential, non-seekable stream. SageMaker’s built-in algorithms like BlazingText and XGBoost support Pipe mode, but custom algorithms often do not unless explicitly coded for it.

Answer

The training instance does not have enough memory

Answer

The IAM role does not have s3:GetObject permission

Answer

The data is compressed and Pipe mode cannot handle compressed data

Answer

The S3 bucket is in a different Region

Question 9

A company uses Amazon SageMaker to train a deep learning model for image classification. The training dataset consists of 500,000 images, each 256x256 pixels, stored in S3. The team uses a single ml.p3.2xlarge instance for training. The training time is unacceptably long (over 48 hours). The team wants to reduce training time without sacrificing model accuracy. They have already optimized the data pipeline by using SageMaker Pipe mode and sharding the S3 dataset. The model is a ResNet-50 implemented in TensorFlow. The team is considering the following options: A) Switch to a ml.p3.16xlarge instance which has 8 GPUs and more memory. B) Implement distributed data parallelism using Horovod across multiple instances. C) Use SageMaker's built-in Hyperparameter Tuning to find optimal hyperparameters. D) Reduce the image resolution to 128x128 to speed up training. Which option will MOST effectively reduce training time while maintaining accuracy?

Accepted Answer

Implement distributed data parallelism using Horovod across multiple instances. Using multiple instances with Horovod for distributed data parallelism can scale training linearly with the number of GPUs, significantly reducing time. A larger single instance (ml.p3.16xlarge) provides 8 GPUs but still limited by single instance. Hyperparameter tuning does not directly reduce training time. Reducing resolution may lose accuracy.

Answer

Switch to a ml.p3.16xlarge instance

Answer

Reduce the image resolution to 128x128

Answer

Use SageMaker's built-in Hyperparameter Tuning

Question 10

A company uses Amazon EMR to run Spark jobs on a transient cluster that processes data from S3. The jobs are failing with 'OutOfMemory' errors. The data engineer has already increased the executor memory. Which additional configuration change would MOST likely resolve the issue?

Accepted Answer

Increase the number of partitions in the data. The 'OutOfMemory' errors in Spark on EMR typically occur when individual partitions hold too much data for the executor's memory to process. Increasing the number of partitions distributes the data more evenly across available memory, reducing the per-partition size and preventing memory overflow during shuffle or aggregation operations. This directly addresses the root cause of memory pressure, whereas simply increasing executor memory may only delay the failure.

Answer

Use fewer, larger instance types for the core nodes

Answer

Increase the driver memory

Answer

Increase the number of executors

Question 11

A data scientist is using Amazon SageMaker to train a model using a custom Docker container. The training job fails with an error message indicating that the container exited with a non-zero code. Which THREE steps should the data scientist take to diagnose the issue? (Choose THREE.)

Accepted Answer

Use the SageMaker Debugger to capture system metrics and output tensors for analysis.. SageMaker Debugger can capture system metrics (e.g., CPU/GPU utilization, memory) and output tensors during training, which helps identify issues like resource exhaustion or problematic gradients that cause non-zero exit codes. It provides deep visibility into the training process without requiring code changes, making it a powerful diagnostic tool for custom container failures.

Answer

Retry the training job with the same configuration; the error might be transient.

Answer

Increase the number of training instances to distribute the workload.

Question 12

A data pipeline uses AWS Lambda to process small files (10-50 MB) from an S3 bucket and write results to DynamoDB. The Lambda function times out after 15 seconds for larger files. The team wants to handle files up to 100 MB without changing the Lambda code. Which solution is MOST cost-effective?

Accepted Answer

Increase the Lambda function timeout to 5 minutes. Increasing the Lambda function timeout to 5 minutes directly addresses the 15-second timeout issue for larger files (up to 100 MB) without requiring any code changes. This is the most cost-effective solution as it avoids additional infrastructure costs (e.g., Glue, ECS, SQS) and leverages Lambda's existing pay-per-execution pricing model, which remains economical for occasional longer-running invocations.

Answer

Use AWS Glue Python shell job to replace Lambda

Answer

Use Amazon ECS with AWS Fargate to run the processing task

Answer

Configure an SQS queue to buffer the S3 events and batch them

Question 13

A data scientist is analyzing a dataset with high multicollinearity. Which TWO techniques can help identify and address multicollinearity?

Accepted Answer

Use Principal Component Analysis (PCA). Correct options: D and E. Variance Inflation Factor (VIF) (E) is a key metric for detecting multicollinearity by measuring how much the variance of a coefficient increases due to collinearity. PCA (D) addresses multicollinearity by transforming correlated features into orthogonal components. Option A is incorrect because a correlation matrix only shows pairwise correlations and may miss higher-order multicollinearity. Option B is incorrect because Lasso regression performs feature selection by shrinking coefficients but does not directly identify multicollinearity. Option C is incorrect because Recursive Feature Elimination (RFE) is a feature selection method that does not detect multicollinearity.

Answer

Plot a correlation matrix

Answer

Apply Lasso regression

Answer

Use Recursive Feature Elimination (RFE)

Question 14

A training job log shows this error. The training instance is an ml.m5.large with 8 GB EBS storage. The training data is 500 MB, and the model size is expected to be 200 MB. What is the most likely cause?

Accepted Answer

The training process is generating large temporary files that fill the instance's local storage. The training job runs on an ml.m5.large instance with only 8 GB of EBS storage. Given that the training data is 500 MB and the model size is 200 MB, there is ample space for those, but the error indicates a lack of storage space. This is most likely caused by temporary files (e.g., checkpoints, logs, or intermediate data) generated during training that fill up the remaining storage. Option A is incorrect because a download issue would show errors like 'Download failed' or 'NoSuchKey', not a storage error. Option B is incorrect because S3 permission issues would result in 'AccessDenied' errors. Option C is incorrect because insufficient RAM would cause a 'MemoryError' or out-of-memory kill, not a storage space error.

Answer

The training data is not fully downloaded from S3 before processing

Answer

The S3 bucket does not have write permissions

Answer

The training instance does not have enough RAM

Question 15

A company wants to build a machine learning model to predict customer churn. The dataset includes customer demographics, usage patterns, and support interactions. The data is stored in Amazon S3. The data scientist needs to perform feature engineering, including creating aggregate features from support interactions and encoding categorical variables. Which AWS service is most suitable for building the feature engineering pipeline?

Accepted Answer

Amazon SageMaker Processing. Amazon SageMaker Processing is the most suitable service because it is purpose-built for data preprocessing and feature engineering within the SageMaker ecosystem. It allows you to run custom Python scripts (e.g., using pandas or PySpark) on managed infrastructure to create aggregate features from support interactions and encode categorical variables, and it integrates seamlessly with SageMaker for model training and deployment.

Answer

AWS Glue

Answer

Amazon EMR

Answer

AWS Batch

Question 16

A data scientist is analyzing a dataset with a continuous target variable and suspects that the relationship between a predictor and the target is non-linear. Which THREE techniques can the scientist use to explore and model this non-linearity?

Accepted Answer

Add polynomial features (e.g., x^2, x^3) and check if model performance improves.. Options C, D, and E are correct. Adding polynomial features (e.g., x^2, x^3) allows a linear model to capture non-linear relationships. Decision tree regressors naturally model non-linear interactions between predictors and the target. A scatter plot with a LOESS smooth curve visually reveals non-linear patterns in the data. Option A (logistic regression) is incorrect because it is for binary classification, not for exploring non-linearity with a continuous target. Option B (Pearson correlation) only measures linear relationships, so it is not suitable for detecting non-linearity.

Answer

Apply logistic regression to binarize the target.

Answer

Compute the Pearson correlation coefficient between the predictor and target.

Question 17

A company is building a near-real-time dashboard using data from multiple sources. They need to aggregate millions of events per second with sub-second latency. The architecture must be fully managed and minimize operational overhead. Which service should they use for the aggregation layer?

Accepted Answer

Amazon Kinesis Data Analytics for Apache Flink.. Amazon Kinesis Data Analytics for Apache Flink is the correct choice because it provides a fully managed, serverless runtime for Apache Flink, which is designed for stateful stream processing at scale. It can aggregate millions of events per second with sub-second latency using exactly-once semantics and built-in checkpointing, meeting the near-real-time dashboard requirements without any infrastructure management.

Answer

AWS Lambda functions triggered by Kinesis Data Streams.

Answer

Amazon EMR with Spark Streaming.

Answer

Amazon Redshift with materialized views refreshed frequently.

Question 18

An e-commerce company uses Amazon Kinesis Data Firehose to deliver clickstream data to Amazon S3. The data arrives at unpredictable rates, with occasional bursts. The company needs to ensure data is delivered within 60 seconds of ingestion, and the data must be partitioned by year/month/day/hour. Which configuration meets these requirements?

Accepted Answer

Set the buffer interval to 60 seconds and enable dynamic partitioning. Setting the buffer interval to 60 seconds ensures data is flushed to Amazon S3 within that time window, meeting the 60-second delivery requirement. Enabling dynamic partitioning allows Firehose to automatically partition data by year/month/day/hour based on the data's timestamp, without needing custom code or additional services.

Answer

Set the buffer size to 1 MB and disable dynamic partitioning

Answer

Use a Lambda function to process data and write to S3 with partitioning

Answer

Use AWS Glue streaming ETL to read from Firehose and write to S3

Question 19

A data engineer is designing a data pipeline that transforms raw JSON files (each 50-200 KB) in Amazon S3 into Parquet format using AWS Glue. The pipeline must minimize data processing costs and handle a high volume of small files (millions per day). The engineer configures a Glue ETL job with Spark, but the job is slow and expensive due to overhead of reading many small files. Which optimization should the engineer implement to reduce cost and improve performance?

Accepted Answer

Use S3 object grouping or batch operations to combine small files before Glue processing.. The primary performance bottleneck with many small files in S3 is the overhead of listing, opening, and reading each file individually in Spark. By grouping or batching small files into larger objects (e.g., using S3 Batch Operations or a pre-processing step), you reduce the number of input splits and task launches, which dramatically lowers the cost and runtime of the Glue ETL job. This directly addresses the root cause of the inefficiency rather than merely scaling resources.

Answer

Increase the worker type to G.2X for more memory per worker.

Answer

Increase the number of DPUs allocated to the Glue job.

Answer

Change the output format from Parquet to CSV to reduce compression overhead.

Question 20

A company runs an e-commerce platform that generates clickstream data in real-time. The data is ingested into Amazon Kinesis Data Streams (100 shards) and processed by AWS Lambda functions, which aggregate data in 1-minute windows and write the results to Amazon S3. The Lambda functions are triggered by the Kinesis stream using the event source mapping. Recently, the company noticed that some records are being processed multiple times, leading to duplicate data in S3. The Lambda function is idempotent, but the duplicates are causing downstream issues. The Lambda function's concurrency limit is 1000, and the batch size is 100. The average processing time per record is 200 ms. What is the most likely cause of the duplicates, and how should it be fixed?

Accepted Answer

Ensure the Lambda function is idempotent and uses the sequence number to deduplicate records.. Lambda functions process records from Kinesis in batches. If the function fails (e.g., due to timeout or error), the entire batch is retried, causing duplicates if some records were already partially processed. To avoid duplicates, the function should be idempotent and should not commit partial results. Option A is wrong because the concurrency is sufficient. Option C is wrong because increasing batch size increases the risk of partial failure. Option D is wrong because a FIFO queue does not integrate with Kinesis.

Answer

Increase the Lambda concurrency limit to 2000 to handle the load.

Answer

Decrease the batch size to 10 to reduce the impact of failures.

Answer

Use Amazon SQS FIFO queue as a buffer between Kinesis and Lambda to guarantee exactly-once processing.